Release IQ3_KS with perplexity info

README.md

@@ -283,6 +283,74 @@ numactl -N 0 -m 0 \
 
 </details>
 
+## IQ3_KS 49.072 GiB (3.816 BPW)
+Final estimate: PPL = 4.7975 +/- 0.02972
+
+<details>
+
+<summary>👈 Secret Recipe</summary>
+
+```bash
+#!/usr/bin/env bash
+
+custom="
+# 47 Repeating Layers [0-46]
+# Note: All ffn_down.* layers are not divisible by 256 so have limited quantization options.
+
+# Attention
+blk\.(0|1)\.attn_q.*=q8_0
+blk\.(0|1)\.attn_k.*=q8_0
+blk\.(0|1)\.attn_v.*=q8_0
+blk\.(0|1)\.attn_output.*=q8_0
+
+blk\..*\.attn_q.*=iq5_ks
+blk\..*\.attn_k.*=iq5_ks
+blk\..*\.attn_v.*=iq5_ks
+blk\..*\.attn_output.*=iq5_ks
+
+# First 1 Dense Layers [0]
+blk\..*\.ffn_down\.weight=q6_0
+blk\..*\.ffn_(gate|up)\.weight=iq5_ks
+
+# Shared Expert Layers [1-46]
+blk\..*\.ffn_down_shexp\.weight=q6_0
+blk\..*\.ffn_(gate|up)_shexp\.weight=iq5_ks
+
+# Routed Experts Layers [1-46]
+blk\.(1)\.ffn_down_exps\.weight=q6_0
+blk\.(1)\.ffn_(gate|up)_exps\.weight=iq5_ks
+
+blk\..*\.ffn_down_exps\.weight=iq4_nl
+blk\..*\.ffn_(gate|up)_exps\.weight=iq3_ks
+
+# Non-Repeating Layers
+token_embd\.weight=iq4_k
+output\.weight=iq6_k
+
+# NextN MTP Layer [46]
+blk\..*\.nextn\.embed_tokens\.weight=iq5_ks
+blk\..*\.nextn\.shared_head_head\.weight=iq5_ks
+blk\..*\.nextn\.eh_proj\.weight=q8_0
+"
+
+custom=$(
+  echo "$custom" | grep -v '^#' | \
+  sed -Ez 's:\n+:,:g;s:,$::;s:^,::'
+)
+
+numactl -N 0 -m 0 \
+./build/bin/llama-quantize \
+    --custom-q "$custom" \
+    --imatrix /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/imatrix-GLM-4.5-Air-BF16.dat \
+    /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/GLM-4.5-Air-128x9.4B-BF16-00001-of-00005.gguf \
+    /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/GLM-4.5-Air-PR624-IQ3_KS.gguf \
+    IQ3_KS \
+    192
+```
+
+</summary>
+
+
 ## IQ2_KL 43.870 GiB (3.411 BPW)
 Final estimate: PPL = 5.0697 +/- 0.03166